This research follows the approach of OMax and Somax. It operates offline and not in real time. Rather than only concatenating audio over time (horizontally), it also concatenates across frequency (vertically), using various tessellation techniques to reconstruct articulation between major spectral components. The result is a tool for compressing and expanding material in both time and frequency; seams are processed along both horizontal and vertical axes. This work is a collaboration with Georges Bloch.
This toolkit is a pipeline that combines continuity-based signal processing with a lightweight probabilistic model to detect, segment, and export vocal events from complex recordings. Beyond detection, it includes a reconstruction workflow based on spectral tiling: each detected “sound element” is treated as a time-frequency tile (or patch) in the MEL spectrum, then selected, arranged, and re-synthesised to build new audio that preserves the character of the source while allowing controlled edits in timing, density, and layering. This makes the project useful for rapid dataset prototyping and reproducible experiment pipelines with visual diagnostics and DAW-ready outputs.